Modeling Subcategorization through Co-occurrence: a Computational Lexical Resource for Italian Verbs

1. Goals and Methodology The aim of this abstract is to introduce LexIt, a freely available lexical resource to characterize Italian verb argument properties in terms of distributional information automatically extracted from large corpora with state-of-the-art computational linguistics methods. Research on automatic extraction of subcategorization frames from corpora has a long tradition in computational linguistics, but to the best of our knowledge this is the first large-scale resource of such type for Italian, aiming at characterizing the predicate valence properties fully on distributional ground. Theoretically grounded on Levin’s assumption that distributional data can be used as “a probe into the elements entering into the lexical representations of word meaning” (Levin, 1993: 14) and methodologically based on the Distributional Semantics framework (Miller and Charles, 1991), our approach proposes to model subcategorization and semantic selection properties through co-occurrence data. Co-occurrence turns out to be a powerful instrument for the study of subcategorization, for more than one reason. First of all, co-occurrences can be automatically extracted from large corpora. Moreover, co-occurrences can used to model the association between verbs and syntactic constructions, arguments and semantic classes as a gradient preference instead of categorical selection. Last but not least, the basic notion of collocation as surface co-occurrence can be integrated with more abstract syntactic or semantic information. We used stochastic association measures (Evert, 2008) traditionally applied to the study of word collocations to evaluate the strength of the correlation between: verbs and syntactic frames, argument slots and the words filling them; argument slots and the semantic classes (or polysemies) selected by them. Currently, LexIt contains more than 3,900 Italian verbs associated with a syntactic and a semantic profile, automatically extracted from La Repubblica Corpus (Baroni et al. 2004). The syntactic profile contains the syntactic frames that best characterize the target verb. The semantic profile is further articulated in two subgroups: the prototypical fillers of each argument slot and the semantic classes abstracted over these fillers.